On Monday, the Chinese artificial intelligence laboratory Z.ai unveiled two new open-source general language models, named GLM-4.5 and GLM-4.5-Air. These models are being marketed as the company’s latest flagship offerings. Both GLM variants are designed as hybrid reasoning models, featuring two operational modes: one focused on complex reasoning and tool use, and the other providing instant responses. Additionally, Z.ai has stated that these models offer agentic capabilities and surpass all existing open-source models globally in performance.
Z.ai Launches Open-Source GLM-4.5 AI Models
In a recent blog post, Z.ai elaborated on the rationale behind its GLM series, emphasizing the aim of developing a comprehensive large language model (LLM) adept at executing a variety of tasks. The company contended that, despite notable strides in generative AI technology, current models from competitors such as Google, OpenAI, and Anthropic fall short of being truly generalistic, as they tend to excel in specific areas while underperforming in others. “GLM-4.5 makes efforts toward the goal of unifying all the different capabilities,” the post asserted.
GLM-4.5 boasts a substantial architecture with 355 billion parameters, including 32 billion active ones, while its counterpart, the Air variant, features 106 billion total parameters and 12 billion active parameters. Both models integrate reasoning, coding, and agentic functionalities into a single structure. They offer a generous context window of 128,000 tokens and include native function-calling capabilities.
For their model architecture, Z.ai implemented a mixture-of-experts (MoE) configuration to enhance computation efficiency during both the training and inference phases. In a departure from approaches like that of DeepSeek-V3, which broaden the MoE layers’ hidden dimensions and expert count, the GLM-4.5 series opts for a narrower but deeper model, believing that increased depth enhances reasoning abilities.
The company also shared innovative techniques utilized in the pre-training and post-training phases within the blog post, aimed at helping the developer community gain insights into the models’ construction.
Performance of GLM-4.5 series AI models
Photo Credit: Z.ai
Z.ai has claimed to have evaluated the GLM-4.5 model against 12 benchmarks, encompassing areas like agentic performance, reasoning, and coding. The firm reported benchmark comparisons with various LLMs from established organizations such as OpenAI, Anthropic, Google, xAI, and Alibaba. According to their internal assessments, the GLM-4.5 ranks third overall, trailing behind OpenAI’s o3 and xAI’s Grok 4 models.
For those interested in exploring these advanced models, Z.ai has made the open weights available on its GitHub repository and on Hugging Face’s listings. Additionally, access to these LLMs is provided through the company’s website and API.
